FAIR compliance, AI-readiness, and ontology maturity of garment datasets: a systematic assessment

This systematic review assessed 64 fashion datasets identified through a PRISMA 2020-guided two-stream search, evaluating each on FAIR compliance, a composite AI-Readiness Score, and a five-level ontology maturity model. The AI-Readiness assessment yielded a grade distribution concentrated in the middle tiers (B: 60.3%, C: 36.5%), while 79.7% of datasets provided only flat attribute annotations (Level 2 or below) on the ontology maturity scale. Beyond diagnosis, the review distills its findings into evidence-derived recommendations for future dataset construction — most consequentially, that new fashion datasets should adopt machine-readable open licenses with persistent identifiers, and should structure annotations at least at the Taxonomy level (Level 3) of the proposed maturity model. Together, the three assessment perspectives — FAIR compliance, AI-readiness, and ontology maturity — and these recommendations provide actionable tools for evaluating future data investments and guiding the development of next-generation fashion AI datasets.

Fashion is one of the most visually rich and commercially significant domains for artificial intelligence (AI) and machine learning (ML) research, sitting at the intersection of computer vision, natural language processing, knowledge representation, and multimodal learning. Clothing items exhibit high intra-class variability, fine-grained attribute distinctions (e.g., “boat neck” versus “square neck”), and complex compositional relationships such as outfit compatibility, making fashion an ideal testbed for advancing AI capabilities in visual understanding, semantic reasoning, and generative modeling (Balchandani et al., 2025; Kouslis et al., 2025). The commercial impetus is equally strong: the global AI in fashion market, valued at approximately USD 1.77 billion in 2025, is projected to reach USD 6.99 billion by 2029 at a compound annual growth rate (CAGR) of 41.3%, and industry analyses anticipate that AI-first fashion companies will capture disproportionate market share in the coming decade (Balchandani et al., 2025; Boston Consulting Group, 2025).

Fashion data also holds broader relevance for AI and engineering research: establishing ground truth for subjective attributes such as “stylish” or “formal” demands specialized ontological frameworks; the convergence of text, image, 3D geometry, and material properties exemplifies the multimodal learning paradigm now dominant across AI; and applications such as virtual try-on, outfit recommendation, and sustainable textile analysis drive innovations with direct societal and economic impact.

Fashion AI research has undergone a paradigm shift over the past decade. Early efforts centered on image classification benchmarks such as Fashion-MNIST (Xiao et al., 2017), conceived as a drop-in replacement for MNIST. Subsequent resources introduced progressively richer annotation: DeepFashion (Liu et al., 2016a) brought large-scale attribute labeling with over 800,000 images; Fashionpedia (Jia et al., 2020) added instance-level segmentation with a structured 27-category taxonomy and 294 fine-grained attributes; and AIpparel (Nakayama et al., 2024) introduced tokenized sewing patterns that represent garments not merely as pixel arrays but as structured, machine-readable design codes. This shift from “image as data” to “pattern as data” fundamentally reconceptualizes what constitutes fashion data for AI systems.

Concurrently, structured knowledge representation has gained increasing recognition. Fashion ontologies such as VetiVoc (Aimé et al., 2016) and the Zalando Fashion Knowledge Graph (Kari, 2018) demonstrate that semantic understanding of garments requires formally structured knowledge systems, and the emergence of large language model (LLM)-driven fashion analysis further underscores the need for ontologically grounded datasets that support semantic reasoning, cross-modal retrieval, and knowledge-augmented generation.

In Korea, substantial government investment has produced resources such as K-Fashion (An et al., 2023) and Korean Hanbok datasets (Hub, 2023), whose nationality-restricted access exemplifies the tension between data sovereignty and open science examined in Section 4.5.

Accordingly, this systematic literature review (SLR) addresses three research questions:

RQ1: What is the current status of open access and accessibility of clothing and fashion databases, evaluated through the FAIR principles (Findability, Accessibility, Interoperability, and Reusability)?

RQ2: How applicable are existing fashion databases to contemporary AI/ML tasks, and what is their AI-readiness level?

RQ3: To what degree do fashion databases incorporate structured knowledge representations, and what is their ontologization potential?

Methodologically, this review adopts the PRISMA 2020 ‘databases, registers and other methods’ design, identifying datasets through two complementary streams: Stream A searches six academic databases for paper-backed datasets, while Stream B applies citation searching and systematic browsing of data portals and repositories to capture datasets distributed without an accompanying academic paper. Every candidate dataset then underwent full-census verification, including live identifier resolution and data-obtainability checks, yielding a final corpus of 64 datasets; the complete selection counts are reported in Section 3.1. The remainder of the paper describes the methodology (Section 2), presents results organized by the three research questions (Section 3), discusses implications, challenges, and future directions (Section 4), and concludes with contributions and limitations (Section 5).

Research design

This study employs a systematic literature review (SLR) methodology following the PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (Page et al., 2021), a framework selected for its established rigor in ensuring transparency, reproducibility, and comprehensiveness. The review covers January 2015 to May 2026, capturing the full trajectory of modern fashion AI research from large-scale datasets to the era of multimodal and generative AI.

Although PRISMA 2020 was developed for reviews of intervention studies, we adopt its core elements — transparent identification, screening, eligibility, and inclusion reporting — and adapt them to dataset cataloging, following dataset-focused systematic reviews in industrial defect detection (Akbas et al., 2024), plant phenotyping (Walsh et al., 2024), and neurosurgical AI (Chan et al., 2025). Items tied to effect-size synthesis (intervention risk-of-bias, meta-analysis, GRADE certainty) are not applicable to resource-cataloging reviews and are omitted, with dataset-specific analogues operationalized through our FAIR compliance, AI-Readiness, and Ontology Maturity frameworks. We considered the PRISMA Extension for Scoping Reviews (Tricco et al., 2018) but adopted PRISMA 2020 for alignment with prior dataset-focused reviews and to support dimensional quantitative scoring. The scope accordingly extends beyond scholarly publications to dataset documentation, competition platforms, and institutional repositories, because many fashion datasets are released without accompanying peer-reviewed papers.

A completed PRISMA 2020 checklist, mapping each of the 27 reporting items to the corresponding manuscript section, is provided as Supplementary material (Section S9). The measured post-revision compliance is 22 of 24 applicable items fully reported (91.7%); 2 items are partially reported (Item #1 — the title describes the work as an AI-Readiness and FAIR-Data assessment rather than using the phrase ‘systematic review’; Item #2 — the abstract does not itemize registration or funding details), and 3 items are justified as not applicable for resource-catalog reviews (Item #12 effect measures, Item #15 certainty assessment, Item #22 certainty of evidence).

PICOD framework for dataset-focused reviews

None of the three dataset-focused SLRs that anchor our PRISMA adaptation (Akbas et al., 2024; Walsh et al., 2024; Chan et al., 2025) reports an explicit PICO-style eligibility framework — a characterization we re-verified against the full text of each paper during revision (Akbas et al. defines eligibility through keyword-based criteria with no occurrence of “PICO”; Walsh et al. uses a keyword-tier scoping methodology with plain inclusion/exclusion criteria; Chan et al. uses plain inclusion/exclusion criteria throughout); we propose PICOD as a standard reporting facet for this review class (facet-by-facet application record: Supplementary Section S10) (see Figure 1).

Two-stream data selection

Following the PRISMA 2020 ‘databases, registers and other methods’ template and the precedent of Lee et al. (2023), dataset identification was organised into two complementary streams. Stream A (databases and registers) queried six academic databases for datasets documented by a peer-reviewed or preprint paper. Stream B (other methods) comprises forward and backward citation searching seeded on the first-round corpus and systematic browsing of data portals and repositories (AI Hub, the Korea Open Data Portal, Kaggle, Hugging Face, Papers With Code, Zenodo, and GitHub) — the identification path for datasets released without an accompanying paper — as reflected in the two-stream PRISMA flow diagram (Figure 2).

Search strategy

The search strategy employed eight keyword groups capturing the multidisciplinary nature of fashion data research — domain terms (fashion/clothing/apparel/garment/textile), resource terms (dataset/database/benchmark/corpus), AI/ML methodology terms, and groups covering ontology/knowledge graphs, virtual try-on/3D garments, recommendation/styling, generation/multimodality, and open access/FAIR principles — combined with Boolean operators (AND, OR) into search strings (full group definitions in Supplementary Table S1).

Database-specific search queries

Six electronic databases — IEEE Xplore, the ACM Digital Library, Scopus, Google Scholar, arXiv, and Web of Science — were searched with database-adapted query strings derived from the eight keyword groups: domain AND resource AND methodology terms, with additional groups as optional database-specific filters, plus language (English, with Korean additionally in Scopus), publication-year (2015–2026), and document-type filters where supported. The full Boolean strings, database interfaces, filters, and per-database result counts are reported in Supplementary Table S1 to ensure reproducibility. Because Google Scholar provides no documented API or fully reproducible query interface, its results were supplemented with targeted forward and backward citation tracking from included studies.

All Stream A searches were executed between May 17 and May 18, 2026, yielding 2,155 records across the six databases. Per-database yields and the complete downstream selection chain are reported in Section 3.1 and the two-stream PRISMA 2020 flow diagram (Figure 2), with query-level documentation in Supplementary Table S1. Datasets released without an accompanying paper were sought through Stream B (Sections 2.2.4 and 2.2.6).

Deduplication and screening procedure

Duplicates were identified by three-pass automated matching — (1) DOI exact match, (2) normalized title exact match, and (3) fuzzy title similarity (difflib.SequenceMatcher threshold ≥ 0.85) — in a custom Python pipeline, with manual comparison where DOIs were unavailable. Of the 489 duplicate records removed from the Stream A search, 193 were removed by DOI match, 213 by normalized-title match, and 83 by fuzzy similarity, yielding 1,666 unique records for title-and-abstract screening. Records identified through other methods are de-duplicated separately within Stream B (Sections 2.2.4 and 2.2.6), with records re-surfaced by the citation network removed as cross-stream duplicates so the two streams remain additive; a consolidated audit trail is provided in the Dataset Collection Validation Transparency workbook (Supplementary material). Two reviewers (YL, Lead; MK, Secondary) independently screened all 1,666 records using Covidence (2024), with disagreements resolved by consensus and a third senior co-author as arbitrator. Inter-rater reliability was assessed using Cohen’s kappa on a random sub-sample of 200 records (12.0% of 1,666; randomly selected using a fixed seed for reproducibility), yielding kappa = 0.82 (95% CI: 0.77–0.87), indicating strong agreement (Landis and Koch, 1977).

Rule-based keyword matching, implemented in two open-source Python tools — auto_screen.py for title-and-abstract screening (Phase A) and auto_screen_phase2.py for full-text screening (Phase C; exclusion codes FT-1 through FT-6) — was used as a reviewer-assistance tool to prioritize records for human evaluation, with full matcher specifications archived in the pipeline artifacts (Section 2.1). Both tools were validated against human-annotated ground truth on stratified random sub-samples (n = 100 and n = 50, respectively; precision, recall, F1, and false-positive rate recorded), and all automated exclusion labels were independently verified by two human reviewers before final eligibility decisions. Automation was not used as the sole basis for any exclusion decision (PRISMA 2020 Item #11).

Because the unit of synthesis was the data resource rather than the paper, records describing the same or derivative datasets were consolidated at the resource level after full-text review under five explicit rules: R1 (successive major versions counted as separate resources), R2 (annotation-only or subset derivatives merged with parent), R3 (multilingual variants merged), R4 (multiple records of one dataset merged into a single primary resource), and R5 (composite datasets counted as new resources); the consolidation log is provided in the Supplementary materials.

Full-text retrieval (Phase B) was attempted for all 186 records passing title-and-abstract screening; inaccessible records (if any) were classified into four sub-categories — no institutional access, retracted (verified against Retraction Watch as of 2026-05-18), broken dataset URL, or unanswered author contact after 30 days — and are reported separately from the 162 records excluded on Phase C content assessment, per PRISMA 2020 Item #16b. The complete retrieval log is available from the corresponding author upon reasonable request.

Exclusion reasons

Across the two screening phases, a total of 1,642 records were excluded from the Stream A database search: 1,480 during title-and-abstract screening and 162 during full-text assessment. Table 1 presents the breakdown of primary exclusion reasons by phase.

Grey literature and supplementary search strategy

In addition to the six academic databases (Section 2.2.1), we conducted systematic supplementary searches across five grey-literature sources — Kaggle, Hugging Face Datasets, Papers With Code, AI Hub Korea, and GitHub, supplemented by the Korea Open Data Portal (data.go.kr) and Zenodo — to capture standalone fashion datasets released without accompanying peer-reviewed papers. All searches were executed on 2026-05-17 using source-adapted queries derived from the eight keyword groups, with 2015–2026 date filters and source-specific quality thresholds (e.g., Hugging Face downloads ≥ 100, GitHub stars ≥ 50). This portal/registry browsing — the second arm of Stream B — produced 6 identification records; two were removed as duplicates before screening (FASCODE, counted once across its two portals, and K-Fashion, already identified and excluded in Stream A), leaving 4 unique standalone dataset records assessed for eligibility, of which 2 were excluded (FASHN Curated — no obtainable research dataset; Korean Hanbok — nationality-gated access) and 2 were included (FASCODE, Second-Hand Fashion). The citation-searching arm of Stream B is reported in Section 2.2.6. The complete search log (queries, filters, dates, result counts) is provided as Supplementary Table S2 in compliance with PRISMA-S Item #4 (online resources and browsing), Item #13 (dates of searches), and Item #15 (total records) (Rethlefsen et al., 2021); these searches address the recurring limitation whereby standalone datasets are missed by academic-database-only strategies (Akbas et al., 2024; Walsh et al., 2024; Chan et al., 2025).

Four-tier source type classification

Stream B — citation searching —Forward and backward citation searches were seeded on the fashion datasets identified in the first round, using Google Scholar and the Semantic Scholar citation graph. The search was executed on 31 May 2026 through the Semantic Scholar Graph API by large-language-model agents under author supervision (see the AI-use statement). The accounting below follows a single chain from units scanned to datasets included, with one distinction stated at the outset: because forward-citation lists were retrieved with per-seed sampling (up to 100 records per seed), the identification totals enumerate citation edges scanned, whereas every subsequent stage counts materialized candidate dataset records. (i) Identification: 16,862 citation edges were scanned — 15,782 forward, 1,069 backward, and 11 from a supplementary sweep of grey-literature registries (per-seed yields in Supplementary Table S3b). (ii) Materialization: screening these edges — removing duplicate edges, records already retrieved by the database search (which the citation network re-surfaces), and non-dataset-introducing publications — materialized 72 candidate dataset records (48 forward, 13 backward, 11 registry); a targeted web search conducted at this stage contributed 10 further candidate records from outside the citation graph, giving 82 candidate records in total (the citation-searching arm of the PRISMA flow therefore comprises 16,872 identification units: 16,862 edges plus the 10 web-search records). (iii) Deduplication: removing 35 duplicates among the 82 left 47 unique candidate records. (iv) Screening and eligibility: of these 47, 45 were assessed for eligibility and 39 new datasets were included — each independently paper-backed (resolving DOI/arXiv), with bibliographic details in Supplementary Tables S3, S8. Stream B thus contributes 39 datasets via citation searching and 2 via portal browsing, and Stream A contributes 23, so the streams are additive without double counting (23 + 41 = 64).

Inclusion and exclusion criteria

Studies and datasets were included under a single, uniformly applied criterion. A resource was eligible if it was (a) a fashion, clothing, or textile dataset with visual and/or textual data; (b) published or released between 2015 and 2026; (c) applicable to AI/ML tasks; (d) supported by verifiable, citable, and persistent documentation — a peer-reviewed paper, a repository or registry record with a persistent identifier, or an archived datasheet; and (e) accessible to all researchers regardless of nationality or geography. Criterion (d) replaces a “peer-reviewed paper required” rule because the unit of synthesis is the data resource rather than the paper; criterion (e) operationalizes international reproducibility, accepting universal subscription paywalls but not nationality- or geography-gated access. Datasets were excluded if they had fewer than 1,000 samples (except unique ontologies, knowledge graphs, or rare modalities), addressed only e-commerce or business analytics without computational methodology, were duplicates or superseded, lacked verifiable persistent documentation, or were nationality-restricted. Applying these criteria uniformly excluded K-Fashion and Korean Hanbok (AI-Hub, Korean-national access; criterion e) and FASHN Curated (a commercial product page with no obtainable research dataset; criterion d), reducing the corpus from 67 verified candidates to 64.

The documentation criterion (d) was operationalized by source type to avoid systematically biasing inclusion against non-traditional datasets: for Type A and Type B records (peer-reviewed papers and preprints with associated datasets), sufficiency was assessed against full-text methodology and supplementary sections; for Type C records (institutional or technical reports, e.g., AI Hub Korea releases, ETRI technical reports, university-affiliated Zenodo deposits), against the institutional metadata schema and curator-published documentation, treating public-institution curatorship as a provenance-assurance pathway equivalent to peer review; and for Type D records (dataset-only releases on Kaggle, Hugging Face, or company pages), sufficiency required completion of at least 12 of the 15 datasheet fields specified by Gebru et al. (2021), supplemented by quantitative engagement metrics demonstrating community validation. The full Type-by-criterion eligibility matrix (Type-stratified) is provided in Supplementary Table S13 (Section S13).

The quality of each included study was assessed using an adapted version of the Critical Appraisal Skills Programme (CASP) checklist for systematic reviews (Critical Appraisal Skills, 2018), supplemented with data-specific criteria covering annotation methodology, documentation completeness, reproducibility of data collection procedures, and benchmark availability. Two reviewers independently assessed quality, with disagreements resolved through consensus discussion. Inter-rater reliability was quantified using weighted Cohen’s kappa across the three rubric dimensions (AS, TAS, OMS) in a full-census independent re-scoring of all 67 candidate datasets (64 included + 3 excluded) by two domain-expert co-authors, with decimal scores binned into four ordinal rubric levels before calculation; the resulting kappa values ranged from 0.65 to 0.72, indicating substantial agreement according to Landis and Koch’s interpretation guidelines. Studies rated as low quality on three or more criteria were excluded from the final synthesis.

Risk-of-bias and reporting-bias assessment

In accordance with PRISMA 2020 Item #14, we systematically assessed five categories of reporting bias relevant to dataset-focused systematic reviews. Because the unit of analysis is a dataset rather than an intervention outcome, conventional funnel plots and Egger’s regression are not applicable; we instead adopted a qualitative-quantitative hybrid framework covering: (i) publication bias (peer-reviewed vs. preprint vs. grey-literature releases, with forward-citation cross-validation for grey-literature-only datasets); (ii) language bias (English / Korean primary coverage; Chinese / Japanese / French / German ecosystems likely under-represented); (iii) database coverage bias (six complementary databases used vs. Chan (Chan et al., 2025)'s single PubMed, with residual CNKI / HCI venue under-coverage acknowledged); (iv) open-access bias (commercial proprietary catalogs excluded by design); and (v) outcome-reporting bias within datasets (13 of the 25 first-round datasets audited for this item lacked baseline model performance on standardized splits). Mitigation efforts included forward/backward citation tracking, explicit supplementary searching of four data registries, and a documented author-contact protocol for request-gated datasets. The dataset-quality rubric (Section 2.3) is complemented by a QUADAS-2-adapted four-domain risk-of-bias tool (Sample Selection, Annotation Process, Data Documentation, Reproducibility) with three applicability concerns (Population / Index / Outcome fit) reported per dataset; the tool was applied with Type-differential evidentiary criteria — multi-annotator agreement and public guidelines for Type A/B, institutional QA documentation for Type C, and datasheet completeness [≥12 of 15 fields (Gebru et al., 2021)] plus engagement proxies for Type D — with the complete Type-by-Domain matrix reported per dataset in Supplementary Table S12 (Section S12).

The data extraction form captured each database’s identity and provenance (name, version, year of publication, originating institution), content (data types, sample and category counts), annotation methodology, schema structure (flat, hierarchical, or ontological), access conditions and licensing, supported AI tasks, benchmark availability, and citation count as a proxy for community adoption. These variables served as input to the three assessment frameworks described below.

Data extraction and analysis framework

Three complementary assessment frameworks were developed to address the research questions systematically. Each framework operationalizes a distinct dimension of database quality (Figure 3).

FAIR-based openness assessment

To address RQ1, we adapted the FAIR guiding principles for scientific data management and stewardship (Wilkinson et al., 2016). Each database was evaluated on four dimensions, each scored on a 4-point scale (1–4): Findability (persistent identifier and rich metadata), Accessibility (availability through standardized protocols), Interoperability (use of formal vocabularies and standards), and Reusability (licensing, provenance, and community standards compliance). Databases were then classified into four access categories: Fully Open (openly downloadable with clear licensing), Academic (available for research with registration), Conditional (restricted by geography, affiliation, or usage constraints), and Closed (commercial-only access). The rubric anchors progress from Score 1 (no persistent identifier, recognized metadata standard, or machine-readable access) through Score 2 (DOI assigned but metadata incomplete or non-standardized) and Score 3 (DOI, rich machine-readable metadata, formal vocabularies) to Score 4 (additionally, a clear open license, community-endorsed vocabularies, and programmatic access through standardized protocols). The rubric was applied independently by two reviewers, with disagreements resolved through consensus discussion.

AI-readiness assessment framework

For RQ2, we developed a composite AI-Readiness Score calculated as a weighted combination of three sub-scores: Annotation Score (AS, weight 0.31), measuring the richness, granularity, and accuracy of annotations; Task Alignment Score (TAS, weight 0.37), evaluating the breadth of AI/ML tasks the dataset supports; and Ontology Maturity Score (OMS, weight 0.32), assessing documentation quality, benchmark availability, and community adoption. Each sub-score was rated on a 1.0–4.0 scale using predefined rubric anchors, ranging from minimal labels to comprehensive instance-level annotation with inter-annotator agreement metrics (AS), from a single task to 6 + tasks with benchmark evaluation protocols (TAS), and from no structured vocabulary to a formal ontology with reasoning support and cross-dataset linking (OMS). The weighted formula is: AI-Readiness = AS × 0.31 + TAS × 0.37 + OMS × 0.32, yielding a score on a 1.0–4.0 scale (Figure 4). TAS carries the highest weight because task alignment directly determines practical utility, while AS and OMS received near-equal weights reflecting their complementary roles in data reliability and semantic interoperability; the robustness of this weighting scheme to alternative specifications is examined in the sensitivity analysis (Section 3.6). We acknowledge that OMS contributes to both the AI-Readiness composite and the independent ontology maturity assessment (RQ3); the overlap is methodologically appropriate because structural knowledge quality is inherent to AI-readiness, but raw correlations between ontology level and AI-Readiness must be interpreted with the partial-correlation caveat in Section 3.6. Databases are then assigned letter grades at fixed thresholds: A (Ready, > = 3.50), immediately applicable to diverse AI tasks; B (Capable, > = 2.50), suitable with minor preprocessing; C (Developing, > = 1.50), requiring significant data augmentation; and D (Deficient, <1.50), insufficient without substantial redevelopment. These boundaries were calibrated by the three co-authors against the observed score distribution to ensure meaningful differentiation. Composite scores are rounded to two decimal places before the grade thresholds are applied; the rounded score is the value reported with the AI-readiness scores in Section 3.3 (e.g., an exact composite of 2.495 is reported as 2.50 and graded B). Note: One ontology-only resource [VetiVoc (Aimé et al., 2016)] is excluded from AI-Readiness scoring as it is not an image-based dataset.

The row-geometric-mean priority method (which coincides with the principal eigenvector for a consistent matrix) yields the priority weights w = (0.307, 0.374, 0.320) for (AS, TAS, OMS), reported to two decimal places as (0.31, 0.37, 0.32) in the scoring formula; the principal eigenvalue λ_max = 3.0001 ≈ n = 3 and the Consistency Ratio CR ≈ 0.000, below the 0.10 acceptability threshold. We caution, however, that for a 3 × 3 matrix the CR is structurally near-zero (a single degree of freedom; RI = 0.58), so a low CR confirms only the transitivity of the three judgments and offers little external validation; the substantive robustness evidence is therefore the sensitivity analysis in Section 3.6, which verifies that grade assignments remain stable across four alternative weighting schemes plus the AHP baseline (Table 2).

Five-level ontology maturity model

For RQ3, we propose a five-level ontology maturity model specifically designed for fashion data, building on the Ontology Spectrum of Daconta et al. (2003) and the W3C Semantic Web Stack (Berners-Lee et al., 2001). The model positions Knowledge Graph (Level 5) above formal Ontology (Level 4) because a queryable knowledge graph subsumes ontology capabilities while adding entity-relation linking and multi-source integration, while acknowledging that some knowledge-graph implementations lack OWL-style formal reasoning support and that the relative positioning may vary by application context. The five levels are: Level 1 (Label), simple categorical labels without structured relationships; Level 2 (Attribute), basic attribute-value annotations; Level 3 (Taxonomy), hierarchical category relationships; Level 4 (Ontology), formal semantic relationships with logical constraints; and Level 5 (Knowledge Graph), full multi-relational entity linking. The model enables a systematic assessment of how well each database’s annotation schema supports semantic AI reasoning.

Literature selection process

The two-stream PRISMA-guided selection process is summarised in Figure 2. In Stream A (databases and registers), 2,155 records were retrieved from the six bibliographic databases (IEEE 459, ACM 283, Scopus 348, Google Scholar 498, arXiv 290, Web of Science 277); after removing 489 duplicate/ineligible records, 1,666 were screened, 186 were assessed in full text, and 23 datasets were included (24 passed full text; K-Fashion was then excluded under the universal-access criterion). In Stream B (other methods), citation searching contributed 16,872 identification units (16,862 citation edges scanned plus 10 targeted web-search records) that reduced, after removal of 16,825 duplicate and non-dataset records, to 47 candidate records and 39 included datasets, and data-portal/repository browsing produced 6 identification records that reduced, after removal of 2 duplicate records, to 4 unique records and 2 included datasets (FASCODE, Second-Hand Fashion). Across both arms, 16,878 records were identified, 16,827 were removed before screening, 51 were screened, 49 were assessed for eligibility, and 41 datasets were included. The two streams yield a final corpus of 64 datasets (23 + 41 = 64). An in-depth FAIR / AI-readiness / ontology assessment (RQ1–RQ3) is reported for all 64 datasets, with scores in Supplementary Tables S17–S19.

The first round (databases and portals) thus contributed 25 verified datasets and citation searching added 39, for a total of 64 (25 + 39 = 64; Figure 2). The 39 newly identified datasets are predominantly recent (29 published in 2023–2026) and span virtual try-on, 3D/4D garment modelling, sewing-pattern generation, multimodal and multilingual fashion understanding, and culturally specific collections (e.g., IndoFashion, AFRIFASHION1600), reducing the geographic and modality gaps of the original corpus.

The 64 included datasets (Supplementary Table S16) span 2015 to 2026 and encompass diverse data types, including 2D images with attributes and segmentation masks, 3D garment models, virtual try-on pairs, outfit compositions, multimodal image-text pairs, dialogue-based coordination data, formal ontologies, knowledge graphs, parametric sewing patterns, and tokenized garment representations. The reviewed databases originate from institutions across North America, East Asia, and Europe, while South Asia, Africa, Latin America, and the Middle East remain underrepresented. 4D-DRESS (Wang et al., 2024) falls below the typical sample threshold (64 garments with 520 motion sequences; the combined count of 584 treats each garment and each motion sequence as a discrete sample) but was included under the rare-modality provision of the inclusion criteria as the only 4D scanning resource identified in the first round.

Across the 64 datasets, publication dates reveal an accelerating trend: only 5 were published before 2018, 3 appeared in 2018, 18 were released between 2019 and 2022, and 38 followed from 2023 to 2026 (see Figure 5 for the timeline of database publication). The post-2022 period also marks a shift from simple classification benchmarks toward complex multimodal and structural resources, including 4D garment data, tokenized sewing patterns, and LLM-driven fashion frameworks.

Each database was evaluated based on its most recent publicly available version as of May 2026. Dataset versioning information, where available, is noted in the Supplementary material.

A complete overview of all 64 included datasets, including identifiers, publication years, sizes, data types, access modes, and source references, is provided in Supplementary Table S5.

FAIR principle assessment (RQ1)

The FAIR-based assessment of all 64 datasets reveals a nuanced picture of fashion data accessibility. The distribution across access categories is as follows: Fully Open, 26.6% (n = 17); Academic, 71.9% (n = 46); Conditional, 1.6% (n = 1); and Closed, 0.0% (n = 0). The dominant access mode is Academic, reflecting the prevalence of research-oriented dataset publication; no included dataset is nationality-gated, as such resources were excluded by criterion (see Figure 6).

Licensing terms among the 46 Academic-access datasets range from CC-BY 4.0, which permits commercial use with attribution, to non-commercial research-only licenses; representative terms are noted in the Supplementary transparency workbook.

A notable qualitative observation — which we term the “openness paradox” — is that several of the most richly annotated, semantically rich resources impose restrictive access conditions; this is an anecdotal pattern among individual high-profile resources rather than a corpus-wide statistical tendency (see the correlation analysis below). K-Fashion (An et al., 2023) — excluded from the corpus under criterion (e) for exactly this reason — features 1.2 million images with 186 attribute categories yet is restricted to Korean nationals through the AI Hub platform, and Alibaba’s FashionAI (Zou et al., 2019) provides hierarchical annotations but limits commercial use, leaving some of the resources most valuable for advanced AI research among the least accessible.

To quantify the relationship between FAIR compliance and access conditions, we computed a Spearman rank correlation (access level coded ordinally: Fully Open = 1, Academic = 2, Conditional = 3) between the FAIR composite score and access level across all 64 datasets. The correlation was strongly negative (ρ = −0.71): more openly accessible datasets tend to score higher on FAIR compliance. This association is partly definitional -- the Accessibility and Reusability rubrics directly reward open access and permissive licensing -- while direct tests of the paradox show essentially no monotonic association between annotation quality or semantic structure and access level (AS vs. access: Spearman ρ = 0.07; OMS vs. access: ρ = −0.08; n = 63 scored datasets), confirming that the paradox concerns individual resources rather than the corpus as a whole.

Per-dataset FAIR scores (Findability, Accessibility, Interoperability, Reusability) for all 64 datasets are provided in Supplementary Table S17.

The FAIR analysis reveals systematic weaknesses in Interoperability (I) and Reusability (R). While most databases achieve adequate Findability through persistent identifiers and descriptive metadata, Interoperability suffers from the lack of standardized fashion vocabularies and cross-dataset schema alignment; even the archetypal open benchmark Fashion-MNIST (Xiao et al., 2017) receives an I1 score, reflecting its absence of structured metadata and standardized attribute vocabularies. Only AIpparel (Nakayama et al., 2024) achieves the maximum FAIR score (F4/A4/I4/R4), benefiting from MIT licensing, a machine-readable tokenized format, standardized vocabulary, and comprehensive documentation (see Figure 7 for the complete FAIR heatmap).

Figure 8 positions the 64 datasets on a four-quadrant map of mean Findability and Accessibility (F + A) versus mean Interoperability and Reusability (I + R), with a threshold of 2.5 on each axis separating FAIR-ready resources from those needing interoperability investment.

AI-readiness assessment (RQ2)

The AI-Readiness assessment of 63 scored datasets (excluding the ontology-only resource VetiVoc) yields a grade distribution of A (Ready, 1.6%, n = 1), B (Capable, 60.3%, n = 38), C (Developing, 36.5%, n = 23), and D (Deficient, 1.6%, n = 1). The concentration at Grades B–C indicates that the majority of fashion datasets are adequate for standard AI tasks but lack the annotation richness, documentation, or benchmark infrastructure needed for advanced research (see Figure 9). The complete per-dataset, grade-grouped listing of all 64 datasets (the 63 scored plus the unscored VetiVoc) is provided in Supplementary Table S15.

Per-dataset AI-readiness component scores (AS, TAS, OMS), composite AI-Score, and grade for all 64 datasets are listed in Supplementary Table S18 (ID-ordered); the same data grouped by AI-readiness grade are given in Supplementary Table S15.

Image classification and attribute recognition

Image classification remains the foundational task for fashion AI. Fashion-MNIST (Xiao et al., 2017), while historically important as an MNIST replacement, receives the lowest AI-Readiness grade (D, 1.19) due to its simplistic 28×28 grayscale format and absence of structured attributes; the grade reflects the framework’s evaluation of fitness for current fashion AI research needs rather than historical importance, though the dataset remains useful as a lightweight prototyping benchmark. In contrast, DeepFashion (Liu et al., 2016a) provides over 800,000 images with 50 categories and 1,000 attributes, achieving a B grade (2.83), and iMaterialist (Guo et al., 2019) advances further with a hierarchical taxonomy of 128 categories.

Recent developments demonstrate the convergence of classification with richer semantic structures. FashionX (Yang et al., 2026), introduced in the OmniFashion framework (a preprint, arXiv:2603.02658), integrates hierarchical annotations with multimodal text-image alignment, achieving a B grade (3.12, below the A threshold of 3.50). K-Fashion (An et al., 2023) provides 186 style attribute categories across 1.2 million images, though its nationality-gated access excluded it from the final corpus (Section 2.3).

Segmentation and parsing

Fashion segmentation and parsing require pixel-level annotation of garment regions and body parts. Fashionpedia (Jia et al., 2020) represents the gold standard, providing 47,461 images with instance-level segmentation masks, a 27-category taxonomy, and 294 fine-grained attributes, yielding the highest AI-Readiness score (A, 3.66) among all evaluated databases. LIP (Look Into Person) (Gong et al., 2017) contributes 30,462 images with 20 human parsing categories and ModaNet (Zheng et al., 2018) offers polygon-based segmentation for street fashion images, achieving C grades of 2.34 and 2.45, respectively. The FASHN Curated resource (Fashn, 2024) extends this lineage with modern parsing annotations but is a commercial product excluded from the review corpus for having no obtainable research dataset (Section 2.3).

3D garment modeling and virtual try-on

The 3D garment modeling category encompasses a spectrum from parametric garment definitions to full 4D capture. TailorNet (Patel et al., 2020) provides 6,170 garment instances with physics-based deformation parameters (C, 2.34); CLOTH3D (Bertiche et al., 2020) generates 118,437 synthetic 3D garments procedurally (C, 1.91); DeepFashion3D (Zhu et al., 2020) contributes 2,078 real-world 3D scans (C, 2.16); and 4D-DRESS (Wang et al., 2024) from ETH Zurich, among the first real-capture 4D garment datasets (64 garments, 520 motion sequences; C, 2.00), represents the frontier of temporal garment modeling.

Virtual try-on has evolved from 2D image warping to more sophisticated approaches. VITON-HD (Choi et al., 2021) provides 13,679 high-resolution image pairs (B, 2.52), while GarmentCodeData (Korosteleva et al., 2025) and AIpparel (Nakayama et al., 2024) represent garments as parametric sewing patterns, enabling direct integration with CAD systems and manufacturing workflows.

This category also reveals a critical scale challenge: 2D datasets routinely contain hundreds of thousands to millions of images, whereas 3D datasets are typically limited to thousands of samples owing to the cost of 3D scanning and modeling. This disparity remains a significant bottleneck for training robust 3D garment models, suggesting that synthetic data generation, as exemplified by CLOTH3D, will play an increasingly important role.

Fashion recommendation and compatibility

Outfit recommendation and compatibility modeling require datasets that capture the relational aesthetics of garment combinations. Polyvore (Vasileva et al., 2018) provides over 164,000 outfit compositions with item-level compatibility annotations (C, 2.45); Fashion IQ (Wu et al., 2021) introduces natural language feedback with 77,684 image pairs and relative captions (B, 2.52); and the Sketchy (LOTS framework) dataset (Girella et al., 2025) pairs sketches with text for multi-conditioned fashion image generation (C, 2.45).

Generative AI and multimodal learning

The generative AI category has experienced rapid growth, driven by advances in diffusion models and large multimodal architectures. Fashion-Gen (Rostamzadeh et al., 2018) provides 325,536 image-text pairs for text-to-fashion generation (B grade, 2.68), while Prompt2Fashion (Argyrou et al., 2024) offers a benchmark for prompt-driven fashion image generation (C grade, 1.91).

Sustainability AI

Sustainability-focused fashion AI is an emerging but underserved category. The Second-Hand Fashion Dataset (Nauman, 2024) provides 31,638 images (Zenodo v3 metadata; 31,997 in the superseded v2 release) of pre-owned clothing with condition and category annotations (C, 2.39), and TextileNet (Zhong et al., 2023) contributes 760,949 textile-material images organized by fibre and fabric taxonomies (B, 2.66). Both remain isolated efforts without the scale or integration needed for comprehensive sustainability AI applications.

Ontology maturity assessment (RQ3)

The five-level ontology maturity assessment reveals that the majority of fashion datasets (79.7%, n = 51) operate at Level 2 (Attribute) or below, indicating that most datasets provide flat attribute annotations without formal relational structure. The distribution across levels is: Level 1 (Label), 4.7% (n = 3); Level 2 (Attribute), 75.0% (n = 48); Level 3 (Taxonomy), 15.6% (n = 10); Level 4 (Ontology), 4.7% (n = 3); and Level 5 (Knowledge Graph), 0.0% (n = 0) (see Figure 10).

At Level 1, Fashion-MNIST (Xiao et al., 2017) assigns each image to one of 10 mutually exclusive categories with no hierarchical structure, and CLOTH3D (Bertiche et al., 2020) and Prompt2Fashion (Argyrou et al., 2024) likewise provide only class-level labels. Level 2 datasets, the majority (n = 48, 75.0%), provide multiple attribute annotations but lack formal relational schemas: DeepFashion (Liu et al., 2016a) annotates images with 50 categories and multiple attributes (texture, fabric, shape, part) as flat lists, and GarmentCodeData (Korosteleva et al., 2025) provides 115,000 + parametric sewing patterns without formalized relationships between pattern components.

Level 3 datasets (n = 10) introduce hierarchical categorization; for example, iMaterialist (Guo et al., 2019) defines a 128-category parent–child taxonomy enabling subsumption reasoning, FashionX (Yang et al., 2026) employs similar hierarchical structures, and TextileNet (Zhong et al., 2023) taxonomically links fiber types to fabric characteristics and sustainability scores. Level 4 (n = 3) represents formal ontology with defined classes, properties, and logical axioms: Fashionpedia (Jia et al., 2020) structures 27 garment categories with 294 part-organized attributes, AIpparel (Nakayama et al., 2024) tokenizes sewing patterns to encode structural relationships among pattern pieces, seams, and construction details, and VetiVoc (Aimé et al., 2016) contributes a standalone modular OWL-based garment vocabulary (see Table 3).

At Level 5 (n = 0), no database in the post-audit corpus achieves the Knowledge Graph designation, which remains a theoretical upper bound characterized by encoding product entities, attributes, and inter-entity relationships in a queryable linked-data format supporting SPARQL queries and external knowledge-base integration (Zamini et al., 2022). The Zalando Fashion Knowledge Graph initiative (Kari, 2018) was excluded through the post-hoc quality audit (Section 4.8) and is referenced only as background literature on industry knowledge-graph efforts.

The strong correlation between ontology maturity and AI-Readiness is noteworthy, though it should be interpreted with caution given that only one dataset achieves Grade A (Fashionpedia (Jia et al., 2020), which also attains Level 4 ontologization); conversely, Level 1 datasets consistently score lower on AI-Readiness. The Spearman rank correlation between ontology maturity level and AI-readiness score was ρ = 0.61 across the 63 scored datasets (excluding VetiVoc, which received an N/A grade), indicating a moderate-to-strong positive monotonic relationship (p < 0.001, 95% CI [0.43, 0.74]). Because OMS constitutes 32% of the composite AI-Readiness score, a degree of definitional circularity is inherent; the non-OMS robustness correlation between the AS+TAS composite and ontology level remains moderate (ρ = 0.49, p < 0.001; see Sections 3.6.3 and 4.7), confirming that the association between annotation quality, task alignment, and ontology maturity extends beyond the definitional overlap.

This correlation has practical implications for dataset creators: organizing annotations taxonomically rather than as flat lists carries a relatively low marginal cost yet is associated with substantially higher AI-readiness scores. We recommend that future fashion datasets adopt at minimum a Level 3 (Taxonomy) structure, with aspiration toward Level 4 (Ontology) for advanced semantic reasoning tasks.

A notable finding across all three assessment dimensions is the emergence of a consistent leading tier: Fashionpedia (Jia et al., 2020), AIpparel (Nakayama et al., 2024), and FashionX (Yang et al., 2026) each combine rich annotations, structured knowledge representation, and reasonable accessibility (see Figure 11 for the radar comparison). Conversely, Fashion-MNIST (Xiao et al., 2017), despite being the most widely cited fashion dataset, consistently ranks at the bottom across all dimensions, highlighting the disconnect between historical popularity and contemporary utility and supporting its treatment as a legacy benchmark appropriate only for pedagogical purposes.

Ontology potential verification

To examine the practical usefulness of the proposed 5-Level Ontology Maturity Model, we introduce the Ontology Feasibility Index (OFI) as an exploratory prioritization metric for assessing each database’s readiness for ontology advancement. The OFI integrates four sub-metrics---the Semantic Richness Score (SRS), the Relationship Feasibility Score (RFS), the Integration Readiness Score (IRS), and the Knowledge Completeness Score (KCS)—via the formula OFI = SRS × 0.25 + RFS × 0.30 + IRS × 0.25 + KCS × 0.20, with RFS weighted highest because relational structure constitutes the core differentiator between maturity levels. Each sub-metric was derived from existing assessment dimensions and capped at 4.00 on a common 1.00–4.00 scale: SRS = AS_score x 0.65 + task_scope × 0.35; RFS = ontology_level × 0.6 + DATA_TYPE_RELATIONAL × 1.5 + OMS x 0.4; IRS = FAIR_I × 0.35 + access_score × 0.30 + OMS × 0.35; and KCS = size_tier × 0.35 + task_scope × 0.30 + data_type_breadth × 0.35. For the knowledge-representation resource without conventional AI scores [VetiVoc (Aimé et al., 2016)], values were estimated from structural properties and FAIR profiles. The Ontology Upgrade Level (OUL) was determined from OFI thresholds: OFI > = 3.50 maps to Level 5, OFI > = 2.80 to Level 4, OFI > = 2.00 to Level 3, and OFI < 2.00 indicates no upgrade.

Per-dataset OFI sub-scores (SRS, RFS, IRS, KCS), composite OFI, and current and predicted ontology levels for all 64 datasets are reported in Supplementary Table S19.

To identify thematic patterns in ontology readiness, all 64 datasets were organized into seven thematic clusters following the AI task categories established in the survey analysis (Table 4). Average OFI varied markedly across clusters: the Knowledge & Pattern Eng. cluster ranked highest (average OFI 3.30), driven by VetiVoc (Aimé et al., 2016) (OFI = 3.62) at Level 4, whereas the 3D/4D & Virtual cluster registered the lowest average OFI of 2.43; the Composition & Recommendation cluster showed the highest average upgrade gap (1.50), indicating substantial latent potential for ontology advancement despite currently immature structural foundations (Figure 12).

At the individual database level (Supplementary Table S19), VetiVoc achieves the highest OFI of 3.62 (SRS = 3.50, RFS = 4.00, IRS = 3.60, KCS = 3.20), followed by AIpparel at 3.44, Fashionpedia at 3.36, FashionX at 3.33, and FashionStylist at 3.26. The Zalando Fashion Knowledge Graph (previously OFI 3.88) was excluded per the post-hoc quality audit (Section 4.8), which explains the change in the corpus-level OFI maximum.

The OFI results exhibit a strong monotonic relationship with AI-Readiness grades: Grade A databases averaged 3.36, Grade B 2.75, Grade C 2.41, and the sole Grade D database (Fashion-MNIST [Xiao et al., 2017)] registered 1.37, indicating that the 5-Level model has prioritization utility for identifying ontology-upgrade potential. Practically, the Composition & Recommendation cluster offers the greatest marginal return on ontology investment, as evidenced by its highest average upgrade gap (1.50), while the 3D/4D & Virtual cluster shows the lowest current average OFI (2.43). Databases with higher FAIR Interoperability scores also achieved higher IRS and OFI values, although this co-movement is partly by construction, as IRS incorporates FAIR-I as a direct input.

Sensitivity analysis

To assess the robustness of the AI-Readiness grading framework, we conducted three complementary sensitivity analyses: (1) weight sensitivity across five weighting schemes (the AHP baseline plus four alternatives), (2) cutoff sensitivity under three threshold configurations, and (3) non-OMS robustness analysis to isolate the contribution of ontology-related metrics to grade assignments. These tests verify that the observed grade distributions are not artifacts of arbitrary parameter choices but reflect consistent underlying patterns in the data.

Weight sensitivity

The default AI-Readiness Score employs weights of w_AS = 0.31, w_TAS = 0.37, and w_OMS = 0.32. We tested four alternative schemes—equal weighting, AS-heavy, TAS-heavy, and OMS-heavy—against this AHP baseline (five configurations in total) to evaluate whether grade assignments are sensitive to the specific weight configuration (Table 5).

Under the default scheme, the grade distribution was A = 1, B = 38, C = 23, D = 1; the corresponding distributions under each alternative scheme are reported in Table 5. Across the four alternative schemes plus the AHP baseline (five configurations in total), at most 8 of the 63 scored datasets changed grade, and every change occurred at the B/C boundary (14 datasets lie within 0.11 of the 2.50 cutoff, default composite scores 2.39–2.52). The framework is therefore robust to moderate perturbations in dimension weights, with grade instability confined to datasets lying near the B/C cutoff.

Cutoff sensitivity

We examined three cutoff configurations for grade boundaries: default (A ≥ 3.50, B ≥ 2.50, C ≥ 1.50), strict (A ≥ 3.75, B ≥ 2.75, C ≥ 1.75), and lenient (A ≥ 3.25, B ≥ 2.25, C ≥ 1.25), which yielded grade distributions of A = 1, B = 38, C = 23, D = 1; A = 0, B = 14, C = 48, D = 1; and A = 2, B = 49, C = 11, D = 1, respectively. Although strict cutoffs shift several datasets downward and lenient cutoffs shift them upward, the ordinal ranking of datasets remains largely preserved, confirming that the framework is not unduly sensitive to the exact placement of grade boundaries.

Partial correlation analysis

To isolate the contribution of ontology-related metrics to the AI-Readiness Score, we computed a partial composite that removes the OMS dimension and renormalizes the remaining AHP weights (AS × 0.456 + TAS × 0.544; i.e., 0.31/0.68 and 0.37/0.68). The Spearman rank correlation between this non-OMS composite and the ontology level, computed with average-rank tie correction, was ρ = 0.49 (n = 63), indicating that ontology maturity captures a moderate share of the variance in AI-readiness even without direct OMS weighting. This supports the convergent validity of the framework and justifies retaining OMS as a dimension that captures variance not fully represented by AS and TAS.

Data quality and ethical considerations

The assessment reveals critical data quality challenges across the fashion database ecosystem. Annotation inconsistency is pervasive: databases use conflicting category definitions, making cross-dataset comparison difficult. For example, what DeepFashion (Liu et al., 2016a) classifies as “T-shirt” may overlap with iMaterialist’s (Guo et al., 2019) “short-sleeve top,” yet no standardized mapping exists between these schemas. The absence of inter-annotator agreement metrics in most datasets compounds this issue, as the subjectivity of fashion attribute labeling is rarely quantified.

Ethical considerations are particularly salient in fashion AI data. Dataset bias takes multiple forms: demographic (over-representation of Western fashion norms and body types), temporal (past trends may not generalize to future styles), and socioeconomic (luxury fashion overrepresented relative to everyday clothing). These biases can propagate through AI systems. The K-Fashion (An et al., 2023) and Korean Hanbok (Hub, 2023) datasets partially address demographic bias through East Asian fashion traditions, but restricted accessibility limits their corrective impact.

Privacy concerns also merit attention. Street fashion photography datasets [e.g., ModaNet (Zheng et al., 2018)] contain images of individuals who may not have provided informed consent for research use. The growing use of AI-generated fashion images [Prompt2Fashion (Argyrou et al., 2024)] offers a partial solution by eliminating privacy risks, though synthetic data introduces its own biases and authenticity concerns.

Beyond bias and privacy, the reviewed datasets raise ethical concerns that extend past FAIR compliance. The cultural and geographic imbalance of the corpus is examined as an external-validity threat in Section 4.7. Several datasets built from web-scraped e-commerce imagery [e.g., DeepFashion (Liu et al., 2016a), Fashion-Gen (Rostamzadeh et al., 2018)] were assembled without explicit consent from the individuals depicted or the original content creators. We urge the adoption of dataset-construction ethics review that goes beyond licensing compliance, including depicted-subject consent, cultural sensitivity review, and transparency about representational limitations.

Standardized fashion ontology establishment

The finding that 79.7% of the 64 datasets operate at Level 2 or below in ontology maturity highlights the urgent need for a standardized fashion ontology. Current annotation practices are ad hoc, with each research group defining its own categories and attributes. This fragmentation impedes data integration, cross-dataset transfer learning, and the development of general-purpose fashion AI models.

A standardized fashion ontology should meet several criteria: (a) hierarchical coverage from high-level garment types to fine-grained attributes; (b) multilingual support reflecting fashion’s global nature; (c) cultural sensitivity accommodating diverse dress traditions; (d) temporal awareness capturing the evolution of fashion categories; and (e) extensibility toward emerging fashion concepts. Existing ontologies such as VetiVoc (Aimé et al., 2016) and Fashionpedia’s taxonomy (Jia et al., 2020) provide valuable starting points, but none achieves comprehensive coverage across all five criteria.

Recent work on LLM-driven fashion ontology construction represents a promising direction, leveraging large language models to automate ontology construction from fashion text corpora. However, LLM-generated ontologies require careful human validation to ensure accuracy and cultural appropriateness.

A methodological caveat concerns the OMS dimension: because OMS contributes 32% of the composite AI-Readiness Score while being correlated with the independent ontology maturity assessment (RQ3), the observed correlation partially reflects definitional overlap. The partial correlation analysis in Section 3.6.3 quantifies this overlap and shows that the relationship persists when OMS is excluded. Future work should develop an ontology evaluation framework fully decoupled from the AI-Readiness composite.

A practical note on framework adoption: the three assessment instruments proposed here can be applied at different levels of completeness depending on the evaluator’s needs and resources. For comprehensive dataset curation or funding decisions, the full triple-framework assessment (FAIR + AI-Readiness + Ontology Maturity) is recommended, as it captures independent quality dimensions that do not always co-vary. For rapid screening or resource-constrained settings, the AI-Readiness grade alone — derived from AS, TAS, and OMS on the 1.0–4.0 rubric — provides a single decision-relevant signal; the sensitivity analysis (Section 3.6) confirms that this signal is robust to moderate weight variation. For ontology investment prioritization specifically, the OFI (Section 3.5) can be applied in isolation. Users need not compute all three frameworks to derive value from the approach.

Multimodal and cross-domain data expansion

The trend toward multimodal fashion data is accelerating: the most recent databases [FashionX (Yang et al., 2026), AIpparel (Nakayama et al., 2024)] combine visual, textual, and structural modalities. However, gaps remain in tactile, auditory, and olfactory modalities (fabric feel, drape behavior, material scent) that matter for comprehensive fashion perception but are not captured by current datasets.

Cross-domain data integration presents another frontier. Fashion data overlaps with domains including retail analytics, cultural heritage, sustainability science, and manufacturing engineering. The integration of fashion data with material science databases [e.g., TextileNet (Zhong et al., 2023) with fiber property databases] could enable AI systems that reason about both aesthetic and physical garment properties.

The emergence of 4D garment data [4D-DRESS (Wang et al., 2024)] signals expansion beyond static images toward temporal and dynamic fashion representation, aligned with the broader AI trend toward video understanding. The integration of garment physics simulation with data-driven approaches is particularly promising, as demonstrated by TailorNet’s (Patel et al., 2020) physics-based deformation combined with learning-based garment modeling.

Sustainability data gap

A critical gap identified by this review is the scarcity of fashion data directed toward sustainability applications. Industry analyses estimate that the fashion industry accounts for 8%–10% of global carbon emissions and approximately 20% of industrial water pollution (Quantis, 2018; Niinimäki et al., 2020), making AI-driven sustainability solutions urgently needed. Yet only two of the 64 datasets [Second-Hand Fashion (Nauman, 2024) and TextileNet (Zhong et al., 2023)] directly address sustainability-relevant tasks, and none provides comprehensive coverage of the fashion lifecycle from raw material through production, use, and disposal.

Future datasets should integrate material composition, manufacturing conditions, transportation footprints, use-phase durability, and end-of-life recyclability data to support AI-driven lifecycle assessment. The intersection of sustainability and fashion AI also invites novel tasks such as automated garment condition assessment for resale, material substitution recommendation, and garment longevity prediction.

Korean fashion AI data infrastructure directions

South Korea has invested substantially in fashion AI data infrastructure via the government-supported AI Hub. The K-Fashion dataset (An et al., 2023) (1.2 million images, 186 attribute categories) is among the largest nationally sponsored fashion datasets, while Korean Hanbok datasets (Hub, 2023) capture traditional costume attributes for cultural heritage preservation.

However, AI Hub’s restriction of data access to Korean nationals creates a significant barrier to international collaboration, limiting the global impact of Korean fashion data and excluding Korean datasets from international validation. We recommend tiered access models: open access for non-commercial research with registration, licensed access for commercial use, and restricted access only for sensitive cultural heritage data, balancing data sovereignty with open science and maximizing the return on national data infrastructure investment.

Korean fashion AI research would also benefit from greater emphasis on ontology development. The K-Fashion dataset’s 186 attributes could anchor a comprehensive Korean fashion ontology formalizing style categories, garment components, and cultural contexts; aligned with international standards such as VetiVoc (Aimé et al., 2016) and Fashionpedia (Jia et al., 2020), it could position Korea as a leader in culturally grounded fashion AI.

Synthesized recommendations for future fashion datasets

Building on the gap analyses in Sections 4.1–4.5, Table 6 synthesizes the principal implications of this review into evidence-derived recommendations for future fashion dataset development. Each recommendation is mapped to the specific finding from which it was inferred; the table functions as a research-implications summary rather than a normative standard.

Threats to validity

We identify four primary threats to the validity of this systematic review. First, regarding internal validity, the OMS component contributes 32% of the composite AI-Readiness Score while also being correlated with the independent Ontology Maturity Level assessment (RQ3), introducing definitional circularity. Partial correlation analysis (Section 3.6) mitigates this concern by demonstrating that the AS+TAS composite (excluding OMS) retains a moderate correlation with ontology level (Spearman rho = 0.49, p < 0.001); future work should fully decouple the two assessments.

Second, regarding construct validity, the AI-Readiness framework weights (0.31/0.37/0.32) were derived through AIJ-AHP over seven literature-derived weighting scenarios (Section 2.4.2), with grade boundaries calibrated by the three co-authors. Sensitivity analysis (Section 3.6) confirms that at most 8 of 63 databases change grade under alternative weight schemes, suggesting reasonable robustness. However, the scoring rubrics involve subjective judgment; we addressed this through dual-independent scoring by two domain-expert co-authors with structured disagreement resolution, verified on the full 67-candidate census (weighted Cohen’s kappa = 0.65–0.72).

Third, regarding external validity, the review is limited to English and Korean publications, potentially excluding relevant Chinese, Japanese, and European-language fashion AI research. The geographic distribution of included datasets is skewed toward Western and East Asian fashion traditions, with underrepresentation of South Asian, African, Middle Eastern, and Latin American fashion data.

Fourth, regarding reliability, Google Scholar does not provide a documented API or fully reproducible query interface. To maximize transparency, we document the exact search date (May 18, 2026, 09:00 KST), browser mode (Chrome incognito), sorting criterion (relevance), and review scope (top 500 results) in the supplementary materials (Supplementary Table S1). Search result screenshots are archived and available from the corresponding author upon request.

Post-hoc quality audit and resource exclusion

Following the initial review, a post-hoc quality audit excluded the Zalando Fashion Knowledge Graph (Kari, 2018), a proprietary internal resource lacking a published name, public download, peer-reviewed documentation, DOI, and stated license; the blog reference is retained as background literature only. A subsequent full-census verification of all 67 candidate datasets applied the universal-access criterion uniformly and excluded three further resources: K-Fashion and Korean Hanbok (AI-Hub access restricted to Korean nationals) and FASHN Curated (a commercial product page with no obtainable research dataset), reducing the corpus to 64 datasets (25 first-round + 39 citation searching); the Korean resources are discussed in Section 4.5 only as exemplars of access barriers, not as included datasets. The five-level ontology maturity model retains Level 5 (Knowledge Graph) as a theoretical upper bound: no included resource meets both its definition and the audit criteria, indicating domain-specific neglect rather than a fundamental obstacle (cf. Bodenreider, 2004; Vrandečić and Krötzsch, 2014).

During a re-audit of the PDF corpus, one file originally filed as ‘DeepFashion’ was found to contain a distinct peer-reviewed dataset paper (Liu et al., 2016b) introducing the Fashion Landmark Dataset (FLD; 123,016 clothing images with 8-landmark annotations). FLD passed the same audit criteria (STRONG, 5.5/6.0) and is retained as a Stream-A dataset alongside the DeepFashion CVPR 2016 entry; all quantitative results in Section 3 reflect the final two-stream corpus of 64 datasets, all assessed in depth.

Three additional borderline inclusions warrant explicit acknowledgment. (i) VITON-HD (Choi et al., 2021) is primarily a virtual try-on method paper, but it releases a high-resolution (1024 × 768) image-pair corpus that has been widely cited as a dataset in subsequent fashion AI research. (ii) AIpparel (Nakayama et al., 2024) is primarily a foundation-model paper; the entry label refers collectively to the model and the accompanying 120,000-garment training dataset that the paper curates and releases. (iii) K-Fashion (An et al., 2023) is the closest available academic citation for the AI-Hub-hosted corpus, which pre-existed the paper. These cases are retained because they correspond to widely used data resources, but they should be read as method- or framework-paired resources rather than canonical dataset publications. Additionally, the dataset released alongside the LOTS framework (Girella et al., 2025) is officially named “Sketchy”; this review uses that name as canonical and cites the LOTS paper as its introducing source.

This systematic literature review has provided a broad assessment of 64 fashion datasets (all assessed in depth) through the lenses of FAIR principles, AI-readiness, and ontology maturity. The key findings are organized by research question.

Regarding RQ1 (FAIR Principles and Accessibility), the central lesson is that the fashion data ecosystem is open in name but not in practice: although no included resource is formally closed, most sit behind academic registration, and FAIR compliance fails precisely on the dimensions that make data machine-actionable — Interoperability and Reusability — where only AIpparel (Nakayama et al., 2024) achieves full compliance. Openness in fashion data currently ends where semantic and licensing machine-readability begin; the openness paradox of Section 3.2 remains an illustrative case-level observation rather than a corpus-wide gradient.

Regarding RQ2 (AI-Readiness), the grade distribution is best read as a diagnosis of where the field’s effort has pooled rather than as a ranking: the overwhelming concentration in Grades B and C (61 of the 63 scored datasets) reveals a mature middle tier of resources adequate for standard supervised tasks, bounded by a nearly empty top tier — only Fashionpedia (Jia et al., 2020) reaches Grade A. The binding constraint on contemporary fashion AI is therefore no longer data volume but annotation depth, documentation, and benchmark infrastructure. We reiterate that these grades index suitability for contemporary human-centered AI applications rather than general dataset quality: the sole Grade D resource, Fashion-MNIST (Xiao et al., 2017), remains a historically foundational benchmark.

Regarding RQ3 (Ontology Maturity), four out of five datasets provide no more than flat attribute annotations (Level 2 or below), and none reaches the knowledge-graph level; combined with the RQ2 result, this locates the field’s deepest deficit in semantic structure rather than in raw data supply. The corpus-level association between semantic structure and AI-readiness — ρ = 0.61 in its raw form, and still ρ = 0.49 once the definitional overlap with the OMS component is removed (Section 3.6.3) — suggests that investment in taxonomic and ontological structure is the single most leveraged improvement available to fashion dataset developers.

The key contributions of this study are threefold:

A proposed AI-Readiness assessment framework (AS × 0.31 + TAS × 0.37 + OMS × 0.32) that provides a standardized, reproducible method for evaluating the suitability of fashion databases for AI/ML tasks.

A five-level ontology maturity model specifically designed for fashion data, offering a systematic framework for assessing semantic knowledge representation capabilities.

A broad, two-stream PRISMA-guided systematic review of 64 datasets that integrates FAIR principles, AI-readiness, and ontologization perspectives, providing a systematic assessment of the fashion data ecosystem that scores each dataset simultaneously against FAIR compliance, AI-readiness, and ontology maturity criteria — a scope that prior fashion AI surveys (Kouslis et al., 2025; Cheng et al., 2022) do not cover, as both are organized by technique or application task and treat datasets descriptively as supporting material rather than as the primary unit of analysis.

Methodological contribution: Towards a PRISMA-Dataset Extension — Beyond the resource catalog, this review advances the methodology of dataset-focused systematic reviews in three ways, which we propose as foundational building blocks for a future PRISMA-Dataset Extension in the spirit of PRISMA-S and PRISMA-DTA: (1) the PICOD framework (Section 2.1.1) adds an explicit Dataset-constraints facet to traditional PICO, addressing a recurring gap observed in Walsh et al. (2024), Akbas et al. (2024), and Chan et al. (2025); (2) the triple-framework quality assessment (FAIR + AI-Readiness + Ontology Maturity, Section 2.4) replaces ill-fitting clinical risk-of-bias tools with a formally validated multidimensional rubric (weighted Cohen’s κ = 0.65–0.72 across the three dimensions, full-census dual scoring of all 67 candidates), supplemented by the QUADAS-2-adapted four-domain RoB tool and the five-category reporting-bias framework of Section 2.3.1, which substitutes for funnel-plot machinery that is mathematically inapplicable to datasets; and (3) the Four-Tier Source Type Classification and Grey Literature Search Strategy (Sections 2.2.4–2.2.5), which capture standalone datasets released without accompanying peer-reviewed papers — as evidenced by the six portal identification records (four unique datasets) in this review — that academic-database-only search strategies systematically miss. Together with PRISMA-S-compliant grey literature search reporting (Item #14) and Type-differential RoB criteria, these mechanisms provide a defensible pathway for including grey literature without compromising evidentiary rigor; the complete grey-literature search log and per-dataset source classification are reported in Supplementary Tables S2, S5. These contributions are submitted for consideration by the PRISMA Steering Committee.

This study has several limitations that should be acknowledged:

The review covers databases published through early 2026; the rapidly evolving nature of fashion AI means that new datasets may have been published after the search period.

The AI-Readiness weights (0.31/0.37/0.32) are literature-derived AHP priors rather than empirically elicited values; as detailed under construct validity in Section 4.7, sensitivity analysis bounds the practical impact of this choice (at most 8 of 63 datasets change grade, all at the B/C boundary), and future work should elicit weights from a multi-stakeholder expert panel, cross-validate them with complementary multi-criteria decision-analysis (MCDA) techniques such as the Best-Worst Method (Rezaei, 2015), and ultimately validate them against data-driven weight learning (Lawrence, 2017; Sambasivan et al., 2021).

The ontology maturity model assesses annotation schema structure rather than the correctness or completeness of individual annotations, which would require dataset-specific quality audits beyond the scope of this review.

The review is limited to English- and Korean-language publications; the resulting coverage limits are discussed as an external-validity threat in Section 4.7.

The corpus is geographically skewed toward Western and East Asian fashion traditions (Section 4.7); culturally specific resources were excluded only when they failed the same documentation, accessibility, or benchmark criteria applied to all candidates.

Future research directions include: developing a community-driven, standardized fashion ontology that integrates the strengths of existing taxonomies; creating FAIR-compliant multimodal benchmarks spanning visual, textual, 3D, and material modalities; establishing sustainability-focused datasets covering the full garment lifecycle; exploring foundation-model pretraining for the fashion domain; and investigating federated architectures that enable cross-database training while respecting data sovereignty. Developing culturally inclusive datasets that represent diverse global fashion traditions remains both an ethical imperative and a scientific opportunity.

The evidence-derived recommendations consolidated in Section 4.6 (Table 6) provide a concrete starting point for translating these findings into future dataset construction, spanning licensing, persistent identifier, annotation, vocabulary, multilingual and culturally inclusive coverage, sustainability metadata, underrepresented modalities, and tiered access dimensions.

As the fashion industry continues its AI transformation, the quality, accessibility, and semantic richness of underlying data resources will be decisive in determining the trajectory of innovation. The AI-Readiness assessment and the five-level ontology maturity model proposed here provide actionable tools for evaluating future data investments, and we encourage the community to adopt, refine, and extend them through collaborative open science initiatives.

The implications of this study extend beyond the fashion domain. The AI-Readiness framework and ontology maturity model can be adapted to other domains requiring data quality assessment, such as medical imaging, autonomous driving, and agricultural AI, although the observed association between structured knowledge representation and AI applicability requires further cross-domain validation. We hope this study serves as both a resource assessment for the fashion AI community and a methodological template for other visually rich domains.